1 /* Optimized memset implementation for PowerPC.
2 Copyright (C) 1997, 1999, 2000, 2003 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
5 The GNU C Library is free software; you can redistribute it and/or
6 modify it under the terms of the GNU Lesser General Public
7 License as published by the Free Software Foundation; either
8 version 2.1 of the License, or (at your option) any later version.
10 The GNU C Library is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 Lesser General Public License for more details.
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; if not, write to the Free
17 Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
24 /* Define a global static that can hold the cache line size. The
25 assumption is that startup code will access the "aux vector" to
26 to obtain the value set by the kernel and store it into this
29 .globl __cache_line_size
30 .lcomm __cache_line_size,4,4
32 /* __ptr_t [r3] memset (__ptr_t s [r3], int c [r4], size_t n [r5]));
35 The memset is done in four sizes: byte (8 bits), word (32 bits),
36 32-byte blocks (256 bits) and __cache_line_size (128, 256, 1024 bits).
37 There is a special case for setting whole cache lines to 0, which
38 takes advantage of the dcbz instruction. */
41 EALIGN (BP_SYM (memset), 5, 1)
44 #define rRTN r3 /* initial value of 1st argument */
45 #if __BOUNDED_POINTERS__
46 # define rMEMP0 r4 /* original value of 1st arg */
47 # define rCHR r5 /* char to set in each byte */
48 # define rLEN r6 /* length of region to set */
49 # define rMEMP r10 /* address at which we are storing */
51 # define rMEMP0 r3 /* original value of 1st arg */
52 # define rCHR r4 /* char to set in each byte */
53 # define rLEN r5 /* length of region to set */
54 # define rMEMP r6 /* address at which we are storing */
56 #define rALIGN r7 /* number of bytes we are setting now (when aligning) */
59 #define rPOS32 r7 /* constant +32 for clearing with dcbz */
60 #define rNEG64 r8 /* constant -64 for clearing with dcbz */
61 #define rNEG32 r9 /* constant -32 for clearing with dcbz */
63 #define rGOT r9 /* Address of the Global Offset Table. */
64 #define rCLS r8 /* Cache line size obtained from static. */
65 #define rCLM r9 /* Cache line size mask to check for cache alignment. */
67 #if __BOUNDED_POINTERS__
69 CHECK_BOUNDS_BOTH_WIDE (rMEMP0, rTMP, rTMP2, rLEN)
71 STORE_RETURN_VALUE (rMEMP0)
72 STORE_RETURN_BOUNDS (rTMP, rTMP2)
76 /* take care of case for size <= 4 */
78 andi. rALIGN, rMEMP0, 3
81 /* align to word boundary */
83 rlwimi rCHR, rCHR, 8, 16, 23
84 beq+ L(aligned) /* 8th instruction from .align */
86 subfic rALIGN, rALIGN, 4
87 add rMEMP, rMEMP, rALIGN
88 sub rLEN, rLEN, rALIGN
92 L(g0): sth rCHR, -2(rMEMP) /* 16th instruction from .align */
93 /* take care of case for size < 31 */
96 rlwimi rCHR, rCHR, 16, 0, 15
98 /* align to cache line boundary... */
99 andi. rALIGN, rMEMP, 0x1C
100 subfic rALIGN, rALIGN, 0x20
103 add rMEMP, rMEMP, rALIGN
104 sub rLEN, rLEN, rALIGN
105 cmplwi cr1, rALIGN, 0x10
109 stwu rCHR, -8(rMEMP2)
110 L(a1): blt cr1, L(a2)
111 stw rCHR, -4(rMEMP2) /* 32nd instruction from .align */
113 stw rCHR, -12(rMEMP2)
114 stwu rCHR, -16(rMEMP2)
115 L(a2): bf 29, L(caligned)
117 /* now aligned to a cache line. */
120 clrrwi. rALIGN, rLEN, 5
121 mtcrf 0x01, rLEN /* 40th instruction from .align */
123 /* Check if we can use the special case for clearing memory using dcbz.
124 This requires that we know the correct cache line size for this
125 processor. Getting the __cache_line_size may require establishing GOT
126 addressability, so branch out of line to set this up. */
127 beq cr1, L(checklinesize)
129 /* Store blocks of 32-bytes (256-bits) starting on a 32-byte boundary.
130 Can't assume that rCHR is zero or that the cache line size is either
131 32-bytes or even known. */
135 beq L(medium) /* we may not actually get to do a full line */
136 clrlwi. rLEN, rLEN, 27
137 add rMEMP, rMEMP, rALIGN
139 bdz L(cloopdone) /* 48th instruction from .align */
141 /* We can't use dcbz here as we don't know the cache line size. We can
142 use "data cache block touch for store", which is safe. */
143 L(c3): dcbtst rNEG64, rMEMP
148 nop /* let 601 fetch last 4 instructions of loop */
150 stw rCHR, -24(rMEMP) /* 56th instruction from .align */
151 nop /* let 601 fetch first 8 instructions of loop */
153 stwu rCHR, -32(rMEMP)
159 stw rCHR, -16(rMEMP) /* 64th instruction from .align */
164 stwu rCHR, -32(rMEMP)
166 add rMEMP, rMEMP, rALIGN
167 b L(medium_tail2) /* 72nd instruction from .align */
171 /* Clear cache lines of memory in 128-byte chunks.
172 This code is optimized for processors with 32-byte cache lines.
173 It is further optimized for the 601 processor, which requires
174 some care in how the code is aligned in the i-cache. */
176 clrlwi rLEN, rLEN, 27
178 srwi. rTMP, rALIGN, 7
182 cmplwi cr1, rLEN, 16 /* 8 */
185 addi rMEMP, rMEMP, 0x20
186 L(z0): li rNEG32, -0x20
190 addi rMEMP, rMEMP, 0x40 /* 16 */
191 L(z1): cmplwi cr5, rLEN, 0
196 addi rMEMP, rMEMP, 0x80
205 /* Memset of 4 bytes or less. */
220 /* Memset of 0-31 bytes. */
225 add rMEMP, rMEMP, rLEN
227 bt- 31, L(medium_31t)
228 bt- 30, L(medium_30t)
230 bt- 29, L(medium_29t)
232 bge- cr1, L(medium_27t)
234 stw rCHR, -4(rMEMP) /* 8th instruction from .align */
240 bf- 30, L(medium_30f)
243 bf- 29, L(medium_29f)
246 blt- cr1, L(medium_27f) /* 16th instruction from .align */
251 stwu rCHR, -16(rMEMP)
262 /* If the remaining length is less the 32 bytes then don't bother getting
263 the cache line size. */
265 /* Establishes GOT addressability so we can load __cache_line_size
266 from static. This value was set from the aux vector during startup. */
267 bl _GLOBAL_OFFSET_TABLE_@local-4
269 lwz rGOT,__cache_line_size@got(rGOT)
273 /* Load __cache_line_size from static. This value was set from the
274 aux vector during startup. */
275 lis rCLS,__cache_line_size@ha
276 /* If the remaining length is less the 32 bytes then don't bother getting
277 the cache line size. */
279 lwz rCLS,__cache_line_size@l(rCLS)
282 /* If the cache line size was not set then goto to L(nondcbz), which is
283 safe for any cache line size. */
287 /* If the cache line size is 32 bytes then goto to L(zloopstart),
288 which is coded specificly for 32-byte lines (and 601). */
290 beq cr1,L(zloopstart)
292 /* Now we know the cache line size and it is not 32-bytes. However
293 we may not yet be aligned to the cache line and may have a partial
294 line to fill. Touch it 1st to fetch the cache line. */
301 blt cr1,L(handletail32)
303 /* We are not aligned to start of a cache line yet. Store 32-byte
304 of data and test again. */
317 /* Now we are aligned to the cache line and can use dcbz. */
320 blt cr1,L(handletail32)
326 /* We are here because; the cache line size was set, it was not
327 32-bytes, and the remainder (rLEN) is now less than the actual cache
328 line size. Set up the preconditions for L(nondcbz) and go there to
329 store the remaining bytes. */
331 clrrwi. rALIGN, rLEN, 5
334 END (BP_SYM (memset))
335 libc_hidden_builtin_def (memset)